On the Complexity of the Monte Carlo Method for Incremental PageRank

TL;DR

This paper demonstrates that the efficiency guarantees of incremental PageRank algorithms depend critically on the assumption of random edge arrivals, showing that adversarial orderings can cause significantly worse performance.

Contribution

It proves that the previously assumed random edge order is necessary by providing adversarial examples that lead to higher running times, extending the understanding of incremental PageRank complexity.

Findings

Adversarial edge orderings can cause higher running times for incremental PageRank.

Random edge arrival assumption is essential for the original efficiency guarantees.

Constructs graphs with worst-case performance under adversarial edge sequences.

Abstract

This note extends the analysis of incremental PageRank in [B. Bahmani, A. Chowdhury, and A. Goel. Fast Incremental and Personalized PageRank. VLDB 2011]. In that work, the authors prove a running time of $O(\frac{nR}{\epsilon^2} \ln(m))$ to keep PageRank updated over $m$ edge arrivals in a graph with $n$ nodes when the algorithm stores $R$ random walks per node and the PageRank teleport probability is $\epsilon$. To prove this running time, they assume that edges arrive in a random order, and leave it to future work to extend their running time guarantees to adversarial edge arrival. In this note, we show that the random edge order assumption is necessary by exhibiting a graph and adversarial edge arrival order in which the running time is $\Omega \left(R n m^{\lg{\frac{3}{2}(1-\epsilon)}}\right)$. More generally, for any integer $d \geq 2$, we construct a graph and adversarial edge order in which the running time is $\Omega \left(R n m^{\log_d(H_d (1-\epsilon))}\right)$, where $H_d$ is the $d$th harmonic number.